Digital Signal Processing Seminar

Automatic target recognition (ATR) encompasses the processing by computer necessary to detect and recognize targets in data acquired from imaging sensors. Examples of such sensors include infrared, multispectral, hyperspectral, laser radar (LADAR), synthetic aperture radar (SAR), and millimeter wave radar. An ATR system is essentially a function that maps images to targets. A major difficulty is that an image is a space of extremely high dimension (equal to the number of pixels) and we must create a mapping that is computationally feasible for applications where recognition has to occur within a few seconds or less. To achieve feasibility, ATR algorithms nearly always extract features that retain the meaningful information while reducing the dimensionality of the space. The recognition problem is then reduced to mapping features to targets. In most ATR systems, the lowest level features are edges. Unfortunately, modern edge detection methods have many defects that hinder ATR performance.

In this report, we presents a new method of edge detection and image enhancement called edge-controlled diffusion and give its extension to vector-valued images. This method overcomes many of the defects of past approaches by locating edges with subpixel resolution, limiting edge distortion, working robustly in the presence of noise, and allowing the merging of edges from all scales. Additionally, it produces an enhanced image suitable for region characterization. Finally, we will propose future research for a multispectral ATR system based on edge-controlled diffusion and part decomposition/synthesis.

A list of digital signal processing seminars is available at from the ECE department Web pages under "Seminars". The Web address for the digital signal processing seminars is http://anchovy.ece.utexas.edu/seminars